Lightning Talk: PT2 Export - A Sound Full Graph Capture Mechanism for PyTorch
summary
Introduces torch.export() as an ahead-of-time full-graph capture for PyTorch that produces an FX graph plus source metadata and static/dynamic shape info, with a guarded contract via shape guards verified at export and compiled into runtime assertions. Uses a tracing backend (dynamo), an Export DB of supported and failing examples, and an API for dynamic symbols (e.g., batch) and run decompositions so backends can target a core operator subset or plug custom kernels.
AI-written summary · about
chapters (9)
- 0:00PT2 Export introduction
- 0:41PyTorch needs whole graph capture
- 3:34Building PT2 Export challenges
- 5:20Soundness in graph capture
- 7:51PT2 Export with static shapes
- 8:56Dynamic shapes for inputs
- 9:37Custom passes and backend integration
- 11:24PT2 Export adoption and future
- 12:49Q&A on export IR
related talks
10mPyTorch 2.x: What’s coming up next for export-path
11mWhat's New in Torch.onnx
08mPyTorch 2.0: TorchDynamo
17mLightning Thunder: Supercharged PyTorch for Modern Hardware
31mPyTorch 2.1 - New Developments
05mPyTorch ONNX Export Support
11mPyTorch 2.0: Dynamic Shapes Support
43mExploring Deep Learning Framework PyTorch